When regulators said AI voices on the phone count as artificial voices

What happened

In January 2024, voters in New Hampshire received robocalls that used an AI-generated voice imitating a US president to discourage them from voting in a primary. The state's attorney general later estimated that the calls numbered between 5,000 and 25,000.

On February 8, 2024, the Federal Communications Commission released a declaratory ruling, adopted unanimously, confirming that calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act (TCPA), a law from 1991 that restricts robocalls and telemarketing. The effect is immediate: callers using AI voices must follow the TCPA's rules, which generally include obtaining the prior consent of the person called, unless an emergency or another exemption applies.

The FCC described the ruling as clarifying and confirming existing rules rather than creating new ones. Law-firm commentary noted that before the ruling there had been arguments that a human-sounding, interactive AI voice might fall outside the "artificial or prerecorded voice" category. The ruling closed that argument.

Why it matters to designers

The ruling is about law, but it sets design requirements for outbound voice assistants:

• Interactivity doesn't remove the rules. A natural-sounding, two-way AI voice is still an artificial voice for these purposes.
• Consent has to be captured and provable. "We have their number" isn't consent.
• Impersonation is a separate harm. Cloning real people's voices for unsolicited calls is exactly what regulators targeted.
• Responsibility is shared. Teams, vendors and platforms may all have obligations, so contracts and controls matter.

Design lessons

1. Make consent a first-class system, with a ledger of who agreed, when, how and to what, checked before every call or message (module 13).
2. Disclose that it's an AI at the start of the call, and never imitate a real person without permission (module 18).
3. Honor opt-outs immediately across all channels (module 13).
4. Respect call windows and frequency limits using the recipient's local time.
5. Log decisions, including why a call was or wasn't made, for audits.
6. Review by market. The same campaign can be compliant in one country and not in another.

How to build it better

A pre-call gate is a small control with a large effect:

Before any outbound call or message, check:
1. Do we have recorded, current consent for this channel and purpose?
2. Is the person on any suppression or opt-out list?
3. Is it within permitted hours in their time zone?
4. Have we reached the frequency cap?
5. Is the script approved, and does it identify us and state that the caller is automated?
If any check fails, do not contact. Log the reason.

The AI opens the call with identification and disclosure, offers an easy way to stop, and hands off to a person when asked.

Try it

Pick an outbound use case, such as appointment reminders. Write the consent wording, the disclosure line at the start of the call, the opt-out path, and the five checks above as specs (module 15) with a test for each failing case.